Related Experiment Video
Updated: Apr 23, 2026
A Cognitive Fusion-guided Prostate Biopsy Using Multiparametric Magnetic Resonance Imaging and Transrectal Ultrasound
Published on: March 21, 2025
Deep-learning computer-aided detection and classification of prostate lesions on biparametric MRI: comparison with
Deepak Jain1, Francisco Restrepo2, Kazuya Yasokawa3,4
1Department of Diagnostic, Molecular and Interventional Radiology, Icahn School of Medicine at Mount Sinai, New York, NY, USA.
Objective:
To assess the performance of a deep learning-based computer-aided detection (DL-CAD) algorithm for prostate lesion detection and classification on biparametric (bp)MRI.
Materials And Methods:
This retrospective, single-center study included men undergoing 3-T MRI of the prostate for suspected prostate cancer (PCa) between July and September of 2022. Using the radiology report as the reference standard, detection performance for high-risk lesions (defined as PI-RADS ≥ 3, 4, 5) by the DL-CAD was evaluated per-patient using sensitivity, specificity, PPV, NPV and AUC; and per-lesion using sensitivity and PPV. Kappa statistics was used to assess per-patient detection and per-lesion classification of PI-RADS ≥ 3 lesions. Clinical and imaging factors associated with discordance between DL-CAD and radiology reports were assessed using Mann-Whitney, Chi-square, and Fisher's exact tests.
Results:
442 adult males (mean age 65 ± 9 years) were assessed. Per-patient sensitivity, specificity, PPV, and NPV for detection of PI-RADS ≥ 4 and 5 lesions were 65.3%/81.2%/62.7%/82.9% and 82.1%/93.8%/65.7%/97.3%, respectively. Per-patient performance for identifying PI-RADS ≥ 3/4/5 lesions was fair-to-excellent: AUC = 0.67 (0.62-0.71)/0.75 (0.71-0.80)/0.92 (0.89-0.96). For detection of PI-RADS ≥ 4 and 5, per-lesion sensitivity was 60.4% and 78.3%, while PPV was 55.0% and 60.3%. Per-patient agreement between DL-CAD and the reference increased with higher PI-RADS scores (kappa = 0.26 (0.18-0.35)/0.46 (0.37-0.55)/0.68 (0.59-0.78)). Agreement on classification of PI-RADS ≥ 3 lesions was moderate (kappa = 0.56 (0.45-0.68)).
Conclusion:
A pre-trained DL-CAD showed good-to-excellent per-patient performance for the detection of PI-RADS ≥ 4 lesions and moderate performance of PI-RADS ≥ 3 lesion classification. Future prospective studies validating the DL algorithm with histopathologic correlation are warranted.
Critical Relevance Statement:
A deep learning computer-aided detection (DL-CAD) algorithm showed good-to-excellent per-patient performance for detection of PI-RADS ≥ 4 lesions, moderate performance of PI-RADS ≥ 3 lesion classification and high negative predictive value, which can be applied in the clinic with knowledge of its limitations.
Key Points:
Clinical validation of deep learning computer-aided detection (DL-CAD) models for the detection and classification of prostate lesions on MRI is urgently needed. A pre-trained DL-CAD algorithm showed fair-to-excellent per-patient performance for detection of prostate lesions on biparametric MRI, with moderate performance for PI-RADS ≥ 3 lesion classification. Identification of false negatives and false positives of prostate cancer detection DL-CAD algorithms is important for future improvement and clinical deployment. A DL-CAD-based prostate cancer detection algorithm with high NPV may reduce interpretation time.
